教育智能体的安全与应用若干问题研究

教育智能体的兴起和教育大数据分析的驱动，让教育人工智能的应用发展欣欣向荣。然而教育智能体面临一系列的特定攻击，安全威胁或事故出现越来越频繁，极可能造成重大事故发生。项目拟对教育智能体的应用及其安全问题展开研究：1）对教育智能体感知信息的敏感数据隐私保护和恶意药饵数据抗干扰研究，不明显降低数据访问效率又保障数据安全；2）基于卷积神经网络的方法，研究如何构建校园安全行为智能分析模型，对校园不良行为判别与预警；3）研究模型对抗性相关技术：基于对抗性样本的可靠鲁棒性和防窃取技术，提高模型的鲁棒性和安全性；4）研究校园安全行为智能分析模型的可解释性和可验证性，增强对模型内部机制以及结果的理解。课题从教育智能体的多模态数据安全、智能分析模型应用、模型安全对抗及解释验证等问题研究出发，形成教育智能体的智能分析模型完整的安全闭环，实现校园安全行为智能分析，尝试为教育智能体的应用及安全供一个良好的解决思路。

The rise of educational intelligence and the analysis of educational big data have enabled the applications of educational artificial intelligence to flourish. However, educational agents are facing a series of specific attacks. Security threats or accidents are becoming more frequent, which is likely to cause serious accidents. The project intends to conduct the application and security issues of educational agents: 1) Research on privacy protection of sensitive data and anti-jamming of malicious bait data by educational agents, which will ensure data security, but not significantly reduce access efficiency; 2) Based on the convolutional neural network method, research how to build an intelligent analysis model of campus security behaviors to identify and warn of bad behaviors on campus; 3) Research on model adversarial related technologies: Reliable robustness and anti-theft technology based on adversarial samples to improve the robustness and security of the model; 4) Study the interpretability and verifiability of the intelligent analysis model of campus security behavior, and enhance the understanding of the internal mechanism of the model and the results. The project research on different problems: multi-modal data security, application of intelligent analysis models, model security confrontation, interpretation and verification, etc. We hope to form a complete closed-loop of intelligent analysis model, and realize intelligent analysis of campus security behavior. We would try to provide a good solution to the application and security of educational agents by this project.